New Algorithm Learns Local Causal Structures with Latent Variables

Identifying the direct causes and effects of a specific target variable from observational data is a crucial task in various scientific and industrial domains. Existing causal discovery methods often either attempt to learn the entire global causal structure, which is computationally intensive, or make unrealistic assumptions about the absence of latent variables and selection bias. This study introduces LoCaLS, a novel algorithm designed for local causal structure learning. It first characterizes a local region within the causal graph that is sufficient for target-specific discovery without needing to reconstruct the full global structure. A theoretical bridge is then established between the causal information derived from this local region and the overall global causal structure. LoCaLS is proven to be sound and complete under standard assumptions, capable of identifying the same direct causes and effects as global methods, but with significantly reduced computational cost. Extensive experiments on both synthetic and real-world datasets, including gene expression data, demonstrate its superior structural accuracy compared to other local methods and its efficiency against state-of-the-art global approaches.